Neuroimaging is a vital tool for brain science in both basic and applied studies — including, for example, studies of cognitive processes and neurodevelopmental trends, and prediction or diagnosis of brain pathology. Despite the advantages of modern imaging technologies, this is still challenging as the data is noisy, high-dimensional, and typically only small sample sizes (as it is expensive to acquire).
Increased access to public neuroimaging datasets has motivated the field to investigate multi-site datasets, which promise an improvement of accuracy rates in the application of advanced computational learning procedures (i.e., machine learning). However, forming a dataset by merely concatenating data from various sites/sources often fails due to batch effects, where the accuracy on a dataset of a model trained on a multi-site dataset is often worse than the accuracy of a model trained on that single site. A promising area for tackling these issues is that of domain adaptation techniques — e.g., transfer learning, which leverages source data to improve related target data performance.
This Research Topic calls for papers focusing on advanced machine learning approaches that can address current challenges in multi-site neuroimaging analysis. Contributions may address homogeneous domain adaptation problems, where the source and target sites have the same modularity of neuroimage data — e.g., multi-site fMRI analysis. Another class of submissions may tackle nonhomogeneous problems, where the source and target sites have different modalities of images. One prevalent use of nonhomogeneous approaches is to improve the quality of low-resolution medical images (such as CT scans) through leveraging high-resolution features (e.g., MRIs). This Research Topic will also cover theoretical studies, which may focus on the development of novel machine learning techniques for multi-site neuroimage analysis — such as probabilistic graphical models, deep learning, multi-view methods, reinforcement learning, etc. Basic and applied studies should indicate successful analyses that relied on advanced domain adaptation techniques to improve the performance of analysis in real-world applications.
Neuroimaging is a vital tool for brain science in both basic and applied studies — including, for example, studies of cognitive processes and neurodevelopmental trends, and prediction or diagnosis of brain pathology. Despite the advantages of modern imaging technologies, this is still challenging as the data is noisy, high-dimensional, and typically only small sample sizes (as it is expensive to acquire).
Increased access to public neuroimaging datasets has motivated the field to investigate multi-site datasets, which promise an improvement of accuracy rates in the application of advanced computational learning procedures (i.e., machine learning). However, forming a dataset by merely concatenating data from various sites/sources often fails due to batch effects, where the accuracy on a dataset of a model trained on a multi-site dataset is often worse than the accuracy of a model trained on that single site. A promising area for tackling these issues is that of domain adaptation techniques — e.g., transfer learning, which leverages source data to improve related target data performance.
This Research Topic calls for papers focusing on advanced machine learning approaches that can address current challenges in multi-site neuroimaging analysis. Contributions may address homogeneous domain adaptation problems, where the source and target sites have the same modularity of neuroimage data — e.g., multi-site fMRI analysis. Another class of submissions may tackle nonhomogeneous problems, where the source and target sites have different modalities of images. One prevalent use of nonhomogeneous approaches is to improve the quality of low-resolution medical images (such as CT scans) through leveraging high-resolution features (e.g., MRIs). This Research Topic will also cover theoretical studies, which may focus on the development of novel machine learning techniques for multi-site neuroimage analysis — such as probabilistic graphical models, deep learning, multi-view methods, reinforcement learning, etc. Basic and applied studies should indicate successful analyses that relied on advanced domain adaptation techniques to improve the performance of analysis in real-world applications.